Can EPro-PnP be used for ego-motion estimation?

[ZhengDonglei]

Thanks for your work and open source code, I am thinking about whether your method can be used for the odometry task of vehicles in the field of autonomous driving.
The pipline is to use ORB to estimate the feature of two adjacent frames of images and calculate their matching relationship.
After obtaining the matching relationship, we use neural network to infer the depth of each image, thereby obtaining the 2D-3D correspondence relationship about the road sign in two adjacent frames of images.
What I want to know is, after getting the 2D-3D correspondence (without w2D), how should I train EPro-PNP to achieve inter-frame ego-motion estimation?